Supertonic 3

Supertonic 3

Turns text into speech in 31 languages, running on your own device through ONNX Runtime with no cloud call. For developers who want fast local voice output with preset voices.

Made by Supertone on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Supertonic 3 is a lightweight text-to-speech system that runs on your own device through ONNX Runtime. Synthesis needs no cloud call. This release covers 31 languages, up from 5 in the earlier version, and the card says it reads more steadily with fewer repeated or skipped words.

Who it suits: App developers, content producers and voice-product teams who want speech generated locally. The open files ship with fixed preset voices, so it fits work that does not depend on a custom voice.

Industry | Who uses it | How they use it | Example
Customer support | Call center teams | Generate spoken lines for a phone flow on their own machines | Turn a greeting into a speech file with a preset voice
Publishing | Audiobook producers | Narrate written chapters in several languages | Convert a passage of a story into a wav file
Games and stories | Writers of character dialogue | Give characters spoken lines in different languages | Voice a short line in Japanese or Korean
Software | Developers of on-device apps | Add speech output without sending text to a cloud service | Call the Python package from an app and save the audio

You can skip it if you need to clone a particular voice from the open files alone. Custom voice styles come from a separate builder that the card mentions.

Where to use it:

• Python on Windows, macOS and Linux
• Fully on your device, with no cloud call for synthesis
• Inside your own scripts and apps through the supertonic Python package
• Hugging Face, where a demo page and the model files live

Use case | Who | What they get back
Call center greetings | Support teams | A spoken line in a chosen voice, saved as audio
Audiobook narration | Publishers and authors | Spoken passages in English or Japanese
Character voices | Game and story makers | Spoken dialogue for characters, including Korean and Japanese
News reading | Media teams | A read-aloud version of a short news text

How to set it up: You need Python and a few minutes for the first download.

• Python on Windows, macOS or Linux
• The supertonic package from pip
• An internet connection on the first run, when the SDK fetches the model assets from Hugging Face

1. Install Python if you do not have it.
2. Install the supertonic package with pip.
3. Run the quick start script below. The first run downloads the model assets.
4. Optionally, download the raw model files with the Hugging Face tool.

pip install supertonic

pip install -U huggingface_hub
hf download Supertone/supertonic-3

How to use it: Create a TTS object, pick a voice style by name, pass your text with a language code, then save the audio.

from supertonic import TTS

tts = TTS(auto_download=True)
style = tts.get_voice_style(voice_name="M1")

text = "A gentle breeze moved through the open window while everyone listened to the story."
wav, duration = tts.synthesize(text, voice_style=style, lang="en")

tts.save_audio(wav, "output.wav")
print(f"Generated {duration:.2f}s of audio")

The card says simple expression tags for a laugh, a breath or a sigh can go inside the text.

Tool | What it does
TTS | Loads the model and downloads its assets on first use
get_voice_style | Picks a preset voice style by its name
synthesize | Turns text into audio for the language code you give
save_audio | Writes the audio to a file such as output.wav

Example: Pass one English sentence to synthesize with the M1 voice style and lang set to en, then save the result as output.wav. You get a wav file, and the script prints how many seconds of audio it generated.

For newcomers and experts: Text to speech turns written words into spoken audio. To try it, install the package, run the quick start script and play output.wav. Change the sentence and the language code to hear other voices and languages.

Under the hood it runs with ONNX Runtime on the local machine, so text never has to leave it. Compared with Supertonic 2, the card reports higher speaker similarity across the shared languages and fewer repeat and skip failures on short and long utterances. Other people have published 41 quantized and 17 fine-tuned versions on Hugging Face.

Text and a language code go in
A preset voice style is chosen
ONNX Runtime synthesizes on the device
Audio comes back as a wav file

Title: Languages supported
Supertonic 2: 5 languages
Supertonic 3: 31 languages

Good to know:

• The open package holds fixed preset voices. Use only voices you have the right to use.
• Custom voice styles are made from reference audio in a separate Voice Builder, and the card says purchased styles are sold there.
• Languages include English, Korean, Japanese, Arabic, Bulgarian, Czech, Danish and German, among the 31.
• License: openrail.
• Cost to run: no API key or account is named for the open files. Voice Builder styles are purchased.
• Data it reaches: the first run downloads model assets from Hugging Face, and the audio file is written on your machine.
• Not for: custom voice cloning from the open files alone.

On Hugging Face, 13 November 2026

• Task: text to speech
• Library: supertonic
• Downloads in the last 30 days: 92,359
• Likes: 1,043
• License: openrail
• Languages: en, ko, ja, ar, bg, cs, da, de
• Quantized versions: 43
• Fine-tunes: 17
• Tags: supertonic, onnx, text-to-speech, speech-synthesis, multilingual, on-device

Links

• Model: https://huggingface.co/Supertone/supertonic-3
• Files: https://huggingface.co/Supertone/supertonic-3/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Supertone/supertonic-3

Then follow the usage in its model card. Nothing in this kit runs by itself.